In recent years, there has been an increasing demand for people's investment behavior through stock investments in order to build individual assets. In stock investment, predicting future market trends is important for investors to reduce their investment risk and to make profits. One of the financial indicators is the volatility index (hereinafter referred to as VI), which represents investors' psychological state toward the market. On the other hand, newspaper media and social media posts represent social conditions and people's psychological states, which may affect the VI. In this study, we predict an increase in the Nikkei 225 VI, the VI in Japan, using newspaper articles and social media posts. Furthermore, in order to verify the usefulness of this study, we conduct a simulation of trading in options using the predicted VI. As a result, we confirmed that the use of both media improved the accuracy of forecasting the rise of the Nikkei 225 VI, and the trading simulation also confirmed its usefulness for profit.

Classification is a central problem in machine learning and requires a classifier. One of the most effective classifiers is a so-called Bayesian network classifier (BNC). Recent studies show that an exact learning of augmented naive Bayes (ANB), which maximizes marginal likelihood (ML) provides higher classification accuracy than any other BNC does. However, maximizing ML has no guarantee to have asymptotic consistency when the true model does not follow a BN. This study proposes a new learning BNC method that asymptotically obtains an I-map with the minimum number of the class variable parameters regardless of whether the true model follows a BN. The proposed method provides more accurate posterior of the class variable than maximizing ML does. Comparison experiments demonstrate the effectiveness of the proposed method.

Long sequence time-series forecasting for counting quantities such as demand, sales, and transactions in stock market is important for various business areas. These kinds of real-world data have properties: such as time dependency, non-linearity, non-Gaussian distribution, zero-inflated and integer values. In this study, we propose a time-series forecasting model for zero-inflated count data. To consider time dependency and obtain long-term outputs, we utilize the Informer which is a long sequence time-series forecasting method based on the Transformer. In addition, we suppose a Poisson distribution and a Bernoulli distribution for the outputs of Informer models to deal with zero-inflated count data properties. We evaluated the method using two artificial and two real-world datasets. The results show that the proposed method can make precise forecasts with long-term adaptation to various trend lines. In particular, the proposed method showed highest prediction accuracy in five of the six experimental conditions using real datasets.

With the recent popularity of e-commerce websites, analyzing customer characteristics from purchase history data has become an important challenge for companies. One effective approach is to describe the characteristics using embedding representations, and “user2vec” is a classic model using neural networks. However, user2vec does not take into account auxiliary information, and the relation between the customer and product is not clearly considered into model learning process. To more accurately capture the characteristics, it is effective to consider auxiliary information, and to evaluate the relationship of products for each customer. In this study, we propose a model that learns both customer and product-specific embedding representations and auxiliary information embedding representations simultaneously, and uses the attention mechanism to associate customers and products. In addition, we perform an evaluation experiment with artificial data assuming purchase history, to demonstrate the effectiveness of the proposed model. Furthermore, we apply the proposed model to actual movie evaluation data, and show a case of customer characteristic analysis.

We introduce a new deal of kernel density estimation using an exponentiated form of kernel density estimators. The density estimator has two hyperparameters flexibly controlling the smoothness of the resulting density. We tune them in a datadriven manner by minimizing an objective function based on the Hyvärinen score to avoid the optimization involving the intractable normalizing constant due to the exponentiation. We show the asymptotic properties of the proposed estimator and emphasize the importance of including the two hyperparameters for flexible density estimation. Our simulation studies and application to income data show that the proposed density estimator is appealing when the underlying density is multi-modal or observations contain outliers.

Multiple robot arms’ path planning for car painting is a process that requires engineers to spend a lot of time using simulators, and there is a need for automation and reduction of planning time. While methods for car welding have been extensively studied, the car painting problem requires specific constraints such as painting a moving car on a line without stopping and painting order to maintain painting quality, making it difficult to directly apply conventional methods. This study proposes a path planning method of multiple robot arms for car painting using evolutionary computation. The proposed method can handle multiple constraints separately in each optimization process according to their characteristics. Experiments of path planning for four arms painting a car side confirmed that the proposed method can find paths similar to those designed by human experts.

Customer segmentation is important for implementing appropriate marketing strategies to meet the different needs of each customer group. The purpose of customer segmentation is to improve the effectiveness of marketing strategies by implementing appropriate measures for each segment, and the formation of similar segments is required to determine the factors that determine the effectiveness of the measures. However, conventional methods do not fully consider this. Therefore, in this study, we propose a method of clustering similar customers based on the impact of feature variables on the effectiveness of measures by using SHAP value vectors, which are known as interpretation methods for machine learning models. This allows us to consider the similarity of the factors that determine the effectiveness of measures, making it possible to implement the most effective measures for each customer segment. We conducted experiments using artificial and actual data to demonstrate the effectiveness of the proposed method.

Many natural phenomena can be modeled as gradient systems, which follow the gradient of a potential function that depends only on the state of the system. Scientists have constructed reduced models of potential functions that reflect the typical features of the phenomena, without contradicting them. However, such modeling requires a large amount of experimental data and specialized knowledge, making it challenging. In this study, we propose a framework to estimate a reduced potential function in a data-driven manner and verify that it is consistent with the phenomenon and contains useful features. First, we estimate the reduced potential function of an unknown physical phenomenon in a data-driven manner using a deep learning model inspired by Hamiltonian Neural Networks (HNN). Then, we try to validate the validity of the obtained reduced potential function and extract useful information to describe the phenomenon. As an example of verifying the usefulness of our proposed framework, we present a case where we apply the proposed framework to numerical calculation data of magnetic domain pattern formation.

In recent years, much attention has been given to deep reinforcement learning, which is one of the artificial intelligence technologies that combines reinforcement learning and deep learning. Deep reinforcement learning, for example, has already shown better performance than humans in games such as Go and Atari video games. Whereas, the progress of its application to real-world tasks beyond artificially limited environments has been slow, and this fact may mean the necessity of other approaches. We focused, in this study, on natural reinforcement learning, which sets an aspiration level and finds quality in rewards. Risk-sensitive Satisficing (RS), an algorithm for natural reinforcement learning, has already demonstrated certain target-oriented exploration and its efficiency in table-based reinforcement learning. However, the current RS employs a Deterministic policy, meaning the difficulty of its application to using probability distributions which deep reinforcement learning draws on. In this study, we extended the Deterministic policy to a Stochastic policy, and verified whether its performances are as good as those of existing table-based reinforcement learning tasks.

The hippocampus is known to be the brain region that replays past experiences. In the context of deep reinforcement learning, experience replay has traditionally been used primarily to improve the sample efficiency of data used to train artificial neural networks and to maintain independence among samples. However, recent advances in neuroscience research have revealed that hippocampal replays occur prior to the onset of locomotion and involve planning that selects the optimal locomotion path from among previously experienced paths, starting from the current location. Inspired by this phenomena, we proposed a mechanism in the Deep Q-Network (DQN) framework to reflect in the current action selection previously experienced paths. This mechanism is described as follows: first, search for trajectories that start from states similar to the current state in the replay buffer that holds previously observed information. Second, reflect the n-step rewards in the past action selections by adding them to the action value of the current state. Our simulation experiments with CliffWalking confirmed that the proposed method allows the agent to maximize returns earlier and to reach the terminal state with fewer steps than normal DQN.

Elastic net, a popular sparse modeling technique, has 2 hyperparameters and, hence, studies on tuning have been conducted. Although the solution map can be approximated with a geometrical shape called Bezier simplex, this requires a high-order polynomial regression, resulting in a complex computation. We thus propose to lower the order by subdividing the Bezier simplex into smaller ones. The subdivision is recursive. Following existing work, we evaluated the method by using qsar-fish-toxicity data. It was implied that the subdivision indeed achieves the same accuracy with a lower order and that parallel computation would reduce the training cost.

Distributed parallel learning is needed due to the growth of deep learning models and datasets. Data parallelization is the easiest distributed learning method to implement, where each GPU has redundant models and batches are distributed. However, as the number of GPUs increases, the batch size increases proportionally and the generalization performance deteriorates due to the loss of the implicit regularization effect of SGD. In this study, we aim to alleviate this large-batch problem by regularizing by the gradient norm.

In order to continuously operate machine learning models, it is necessary to find the causes of prediction errors and take appropriate measures. At this time, it is possible to estimate the effectiveness of measures for each cause by quantitatively evaluating the impact of each cause on the prediction error. This study proposes a method to decompose the prediction errors occurring in operation into contributions from multiple causes. We verified the usefulness of the proposed method through experiments.

Global optimization of tensor low-rank decomposition by reducing ranks is fundamentally difficult due to the non-convexity of the associated cost function. Instead of ranks, we use interactions between modes as an alternative concept of ranks and formulate convex tensor decomposition that controls mode interactions by minimizing the Kullback--Leibler divergence from input.

Multi-Objective Reinforcement Learning (MORL) is a generalization of standard reinforcement learning that aims to balance multiple, possibly conflicting, objectives. A common challenge in MORL is to learn policies that correspond to any Pareto optimal solution, especially when the Pareto front is non-convex. In this paper, we propose a novel method that learns a single policy that directly optimizes the hypervolume metric, which measures the volume dominated by a set of points in the objective space. The main idea is to transform the multiple objective values into hypervolumes and apply Watkins' Q-learning algorithm to learn a policy that maximizes the hypervolume. Moreover, our method can adapt the policy to achieve any desired Pareto solution without retraining. We call our method hypervolume maximization Q-learning, and present two variants of it–a tabular version and a deep learning version. We evaluated our method on the Deep Sea Treasure benchmark, a non-convex MORL problem, and show that it can effectively learn policies that achieve all Pareto solutions.

We present a system that predicts the types of formulas in a math text using a neural language model and type inference. Firstly, we enumerate possible types of a formula with type inference. Since a formula generally has multiple interpretations, we cannot fully determine its type without the context. Secondly, we input the formulas and the context into a neural language model and predict their types with certainty scores. Finally, we select the type which obtained the highest score for each formula. Experimental results on a math problem dataset show that, unfortunately, the accuracy of the prediction deteriorates when we combine symbolic type inference with statistical prediction.

In the natural language processing in the financial domain, many studies focus on the analysis of documents at a certain point in time, while various documents such as financial statements and financial results are published regularly.For investors with many stocks, it is not easy to read every detail of documents that exist at two points in time about each company continuously.It is also difficult to find out what has changed since the last time published.In this study, we propose a task to extract differences from two similar sentences in the two financial results written about the same company.We automatically extract similar parts from the two documents.For the extracted parts, we manually extract the differences between the two sentences.In addition, we conduct an evaluation experiment with a pre-trained language model.

In research and development, previous works are useful for surveying technical methods. However, in the recent information explosion era, the number of papers that individual researchers can carefully read is limited, and it is difficult to comprehensively collect information. Recently, open-access archive (such as arXiv.org) for academic articles have been paid attention, and have become useful knowledge sources both in quality and quantity. Based on the knowledge source, it would be very useful if technical information such as ideas and algorithms could be referred to from a large number of research in specific fields. Therefore, in this research, we proposed an information recommendation technique that recommends reference information for target problems using paper contents as a knowledge source. As a result, in experimental evaluation with a large-scale paper archive, we confirmed the effectiveness of the proposed method.

Citation count is the most recognized metric for measuring the impact of scholarly papers. However, cited papers contribute to a citing paper in a variety of ways. In order to accurately measure the individual impact of cited papers, it is crucial to quantify their respective degree of contribution to a citing paper. This study proposes a method to automatically identify which papers are highly meaningful to a citing paper by extracting the description of a cited paper from a citing sentence for use in identifying meaningful references. Experimental results demonstrated the effectiveness of the extracted description.

In recent years, there has been a lot of research focused on using language models instead of knowledge bases. Language models have many advantages compared to structured knowledge bases, such as not requiring manual definition of information attributes and relationships and being able to search more data in a more flexible and efficient manner. However, their performance is still developing, and there are still many hurdles to overcome, such as the inability to predict compound nouns. This study specifically focused on the knowledge of specialized compound nouns related to chemistry and investigated how accurately knowledge in a specific field could be extracted. Specifically, by using SciFive, which was further trained with T5 on biomedical papers, and by performing additional training on abstract data contained in Scopus, the study aimed to improve the accuracy of extracting specialized knowledge in chemistry. The results confirmed how accuracy changes depending on the amount of data used for additional training, with a decrease in accuracy with less data and an improvement in accuracy with relatively more data. These results demonstrate further potential for attempts to extract knowledge from language models.

Inference between different modalities has been actively studied in recent years. We focus on Visual-textual Entailment (VTE), one of the most critical tasks for multimodal inference. A variety of deep learning-based approaches have been proposed for the VTE task, but they have difficulty in accurately handling numerals. In contrast, approaches based on logical inference can successfully deal with numerals. However, since the previous logic-based approaches use automated theorem provers, their computational cost significantly increases for problems involving many entities. In this paper, we propose a logic-based VTE system with model checking and knowledge injection. We create a dataset for the VTE task containing numerals and negation to evaluate the extent to which VTE systems correctly understand those phenomena. Using this dataset, we show that our system solves the VTE task with numerals and negation more robustly than the previous approaches.

Natural Language Inference (NLI) tasks that require temporal inference remain challenging for pre-trained language models (LMs). Although various datasets have been created for this task, they primarily focus on English and do not address the need for resources in other languages. In this paper, we present a Japanese NLI benchmark for temporal inference. To begin the data annotation process, we create inference templates consisting of various inference patterns based on the formal semantics test suites. We then automatically generate diverse NLI examples by assigning nouns, verbs, and temporal expressions to the templates using the Japanese case frame dictionary. We evaluate the generalization capacities of monolingual/multilingual LMs by using controlled splits of our dataset. Our findings demonstrate that LMs struggle with handling specific linguistic phenomena such as habituality.

This study aims to develop a method to predict physical events using language. It focuses on collisions between objects and takes into account their characteristics such as mass, speed, and environmental factors. Based on the CLEVRER dataset, this study creates a natural language dataset describing causal relationships in the physical environment. This dataset is used to infer the post-collision state of objects and express these predictions as natural language sentences. The study demonstrates that it is possible to predict events in the physical world using language, and to express these predictions as inferences in the symbolic world.

Human reasoning is not only based on logical implication relations, but also on common sense inference based on everyday knowledge. In this study, we consider knowledge in this case as natural language sentences themselves, which can express complex contents, and use Clause Patterns to learn natural language inferences in which both premises and consequences are sentences directly from a corpus as a deep learning model. We pre-processed 1.57 million premise/consequence pair sentences extracted from the corpus, trained on those pair sentences for which the inference is plausible using T5, and constructed a non-logical linguistic inference by natural language sentence generation. On the test data, we manually evaluated the inferences generated from the premises, and found that 65.8% of the inferences were valid. We further discuss the reasons why valid inferences were not obtained and discuss possible improvements and future possibilities.

Recognizing Textual Entailment (RTE) is an important task, which is applied to question-answering and machine translation. One of the main challenges in logic-based approaches to this task is the lack of background knowledge. This study proposes a logical inference system with phrasal knowledge by comparing their visual representations based on the intuition that visual representations facilitate humans to judge entailment relations. First, we obtain candidate phrase pairs for phrasal knowledge from the process of logical inference. Second, using a Vision-and-Language model, the visual representations of these phrases are acquired in the form of images or embedding vectors. Finally, the obtained visual representations are compared to determine whether to inject the knowledge corresponding to the candidate or not. Besides simple similarity between phrases, asymmetric relations are considered in comparing visual representations. Our logical inference system improved the accuracy on the SICK dataset compared with a previous logical inference system, SPSA.

To obtain a genuine meaning for a natural language sentence, it is necessary to understand the connection between words or phrases in a language and various kinds of real-world information. One of such real-world information might be odors. Previous studies investigated whether word embeddings from word2vec can acquire odor information. However, their model, trained with general corpora, does not have much odor information due to a small volume of corpora related to odors. In this paper, we propose TOLE, Thesaurus-enhanced Odor-adaptive Linguistic Embeddings. TOLE retains the odor information with domain adaptation and word-level contrastive learning on pre-trained language models. As a result, TOLE can improve the similarity between odor embeddings from odor descriptors and linguistic embeddings.

With the spread of social networking services (SNS) in recent years, opportunities for text-based communication have increased. However, there are cases in which awkward exchanges continue, especially with people who are not close to each other. In this study, we propose a support system to insert a simple “aisatsu” at the beginning of a reply to a comment made by the other party. This will facilitate the exchange of dialogues and improve the impression of the dialogues.

We are developing a novel system for evaluating and providing feedback on job interview responses. In order for the system to give better feedback, it is crucial to capture the discourse structure of interview responses. Towards creating a corpus of job interviews annotated with discourse structure, we explore the PREP method, which is widely recognized as an effective discourse structure for achieving high scores in job interviews. We conduct trial annotation on 220 job interview responses and discuss the challenges and limitations of the PREP-based annotation scheme. Finally, we investigate the relationship between the PREP structure and the interviewer's evaluation by analyzing annotated interviews.

Shizuoka is located on the border that divides the dialects of Japan into East and West, and the Shizuoka dialect is notable for having a variety of vocabulary and expressions not found in Standard Japanese. Shizuoka dialect uses multiple inferred expressions such as tsura, zura, dara, and ra. The purpose of this study is to conduct a large-scale survey on the use and understanding of Shizuoka dialects, and to estimate what kind of changes have occurred in the past, and predict future change trends by using the analysis methods of Multinomial Regime Switching Model, and Multi-category Order Statistics. The results of this study show that the inferred expression tsura has almost disappeared and that zura has also shown a marked declining trend. The results also suggest that these changes will continue in the future. Furthermore, the use of dara, the inferential expression based on objective evidence, was found to be more common among middle-aged adults. On the other hand, the use of dialectal forms declined among younger age groups, and the same for the inferential expression ra based on speaker judgment.

We describe a method for generating and singing a dodoitsu that summarizes a newspaper article. Specifically, we generate a training data set by modifying newspaper headlines to match the form of a dodoitsu, and train T5 to automatically generate a dodoitsu from an article. As a result, we were able to generate a dodoitsu for about 96% of the articles, and its quality as a news summary was about the same in ROUGE-1 precision and about half in recall compared to the headlines automatically generated by T5. In addition, by modifying DiffSinger, we succeeded in synthesizing a dodoitsu singing voice whose long tones are more stable than the original DiffSinger.

Several dialogue system studies have attempted to replicate a specific speaker desired by users. Typically, to assess the ``speakerness'' of a specific speaker subjectively, evaluators should be familiar with the target speaker. However, when using a corpus collected from the web and crowdsourcing, it becomes challenging to find evaluators familiar with the target speaker. To address this issue, we propose a novel method for assessing target speakerness through dialogue comparison that can be utilized by non-acquainted evaluators. We evaluate the effectiveness of this method using both expert annotators and non-expert crowdworkers, discussing its validity as a subjective evaluation tool for speakerness. Additionally, we train and examine the performance of a baseline model for assessment of target speakerness through dialogue comparison.

To prevent epidemic spread in organizations, it is important to control contact opportunities. Since the epidemic situation changes day by day, the fixed suppression of contact opportunities may become an excessive measure and lead to business opportunity losses. If organizations can plan contact opportunities on timeline, they will take necessary measures against epidemic spread while preventing the opportunity losses. Therefore, we propose a method to plan contact opportunities on timeline to prevent epidemic spread according to the characteristics of individual organizations. In our proposing method, a timeline of contact intensities that enables organizations to achieve the number of acceptable infected people is explored coupling an agent-based epidemic model with an inverse analysis. Simulation experiment reveals that it is possible to obtain an optimized timeline of contact intensities every certain periods. The proposing method will be useful for organizations to plan epidemiological measures like social distance based on several periods.

Visual information such as evacuation signs has been commonly used to support efficient evacuation in indoor evacuations. Recently, evacuation guidance systems with dynamic evacuation signs have been discussed, aiming to adapt to the dynamically changing evacuation environment and mitigate congestion that occurs during evacuation. Considering the severe environment during disaster evacuation, these systems must be robust; it is undesirable that the entire system is significantly affected by the malfunction of some components. A decentralized architecture, in which each evacuation sign functions independently without a central control mechanism, is considered suitable for these systems. We assume a system consisting of evacuation signs and sensors that function independently and examine the effectiveness of such a system in crowd evacuation. The analysis employing evacuation agents with herd behavior revealed that the system could decrease evacuation time by reducing congestion during evacuation compared to a system with static signs. Furthermore, the system can support efficient evacuation for evacuation routes that change during evacuation.

Diffusion mechanism design is a new research paradigm in the literature of mechanism design, which aims to incentivise agents to invite as many colleagues as possible to participate in a mechanism. In this paper we apply diffusion mechanism design to the facility location game, one of the most well-studied model in MAS domain. We first provide a general impossibility result on the existence of possibly randomized facility location mechanisms that are strategy-proof, Pareto efficient and fully anonymous. We then present two strategy-proof mechanisms that satisfy some weaker notions of anonymity.

This study aims to estimate the number of backup workers required to continue production when the infection parameters change to various situations. Enough workers are necessary to run the production site. However, it is difficult to estimate the number of backup workers because of the complexity of the production site and the sudden absence of workers by infectious. In this paper, we used an Agent-based modeling production site infection simulator to compare changes in the number of backup workers under the following conditions: two virus types, three mask-wearing situations, and two backup target ranges. The results show the infection parameters including the duration of each symptom period are important to estimate the number of backup workers, the backup workers with masks need to be available for 10-25% of the production site, and the backup target range should be determined in combination with measures to reduce the probability of infection.

In recent years, personal mobility vehicles, such as Segway and electric scooter, have attracted attention as a new means of transportation. Interactions between riders and pedestrians should be understood to evaluate the pedestrian comfort and safety. Previous studies modeled pedestrian, Segway and electric scooter movements using the Social Force Model. In this study, we improved the model for simulating pedestrian and electric scooter. Furthermore, we conducted mixed traffic simulations of pedestrians, Segways and electric scooters. The simulation results show that personal mobility vehicles move differently at high and low pedestrian density.

Automated negotiation is a technology that searches for agreeable conclusions where agents with different interests and objectives negotiate with each other. The purpose of this technology is to substitute or support human-to-human negotiation. Conventional automated negotiation models assume that all negotiators are sure to fulfill their agreements if they reached agreements at the end of the negotiation. In our society, however, it's often uncertain whether the agreements will be fulfilled. To solve this, we propose a negotiation model in which agents are not necessarily certain to fulfill the agreements. Furthermore, we propose a meta-strategy that changes the negotiation strategy according to the observed results of the other party's non-performance. Simulation results show that the proposed meta-strategy obtained higher utility values than the Boulware (Conceder) strategy. From this research, we can say that it's important to choose negotiation strategy in consideration of the previously observed performance results of other agents.

In a multi-agent environment, agents should often be required to choose cooperative behavior to solve tasks. The agents usually owes such behavior to special rules designed by humans, especially in an environment consisting of heterogeneous agents, but it is impossible to design such rules for myriad situations. Thus, it has been proposed to learn such cooperative behavior with reinforcement learning in such a hetero-agent environment where they have to collect targets. That method, however, highly depends on the environment; the learned policies do not work in other environments at all, even in similar ones. This work alleviates the problem by defining the state space relatively, i.e., the state space is defined by the relation between the agents and the target. The experimental results show that the policies obtained by the proposed method work well in other, similar environments, as well as in the identical one.

The sharing economy is a new type of economic system in which assets and services are shared between private individuals. Ride-sharing is one of the services for sharing economy, and is the transportation service that connects car owners with users who want to use a car. Taxi-Sharing is one of the representative ride-sharing services. In this paper, the main focus is laid on the Taxi-Sharing problem based on Coalition Structure Generation with Services (TSPSCSG). A formal framework is defined and some decision and optimization problems for TSPSCSG are provided. In the experiments, the TSPSCSG is solved by using real data on taxi fares in the Hanshin area.

This paper examines the impact of trust and betrayal on networks. A generative model of a network of trust is proposed that introduces a mechanism of trust and betrayal. Assuming that a node is a person in the network and an edge is a commitment relationship between two persons, the proposed model generates a network by linking edges, which are commitment relationships, when two persons in the network have an opportunity to trade and trust each other. The results show that betrayal strengthens the ties in the network, but on the other hand increases the number of isolated nodes.

Mechanism design, a branch of economics, aims at designing rules that can autonomously achieve desired outcomes in resource allocation and public decision making. The research on mechanism design using machine learning is called automated mechanism design or mechanism learning. In our research, we construct a new network based on the existing method for single auctions and aim to automatically design an mechanism by applying it to double auctions. Especially, we focus on the following four desirable properties for the mechanism: individual rationality, balanced budget, Pareto efficiency, and incentive compatibility. We conducted experiments assuming a small-scale double auction and clarified the convergence of the obtained mechanism. We also confirmed how much the learnt mechanism satisfies the four properties compared to two representative protocols. As a result, we verified that the mechanism is more budget-balanced than VCG protocol and more Pareto-efficient than MD protocol, with the incentive compatibility mostly guaranteed.

This paper aims to develop an automated negotiation meta-strategy and proposes an approach that automatically selects a strategy according to the opponent from a set of multiple strategies using clustering. The proposed method makes groups of the strategies of possible negotiation opponents. It learns an effective bidding strategy corresponding to the representative point of each cluster using deep reinforcement learning, which is for the average agent in each cluster and is strong on average against the agents in the cluster. We analyzed the number of clusters of the strategies retained by the proposed method and found that the individual utility tends to be higher when the number of clusters is small, and especially the utility was highest when the number of clusters is 3. In addition, the negotiation simulation experiments demonstrated that the proposed method gained higher individual utility than the previous study.

We have focused on the consolidated transportation, which is particularly difficult to control cargo volume, among truck transportation. Consolidation transportation has problems with vehicle arrangements and profitability. In this paper, we study a method to control the fluctuation of cargo volume by determining the transportation unit price from the expected cargo volume. We show that dynamic pricing changes the behavior of shippers, reducing fluctuations in cargo volume by simulation. We also show that it may increase profits of the logistics companies without increasing shipping costs paid by shippers under the conditions of our simulation.

We report results of an experimental evolution to study the mechanism of music style evolution. Creative cultures including music are developed by descent-with-modification processes of knowledge about creating complex artifacts, but how new creation styles emerge and the influence of listeners/consumers through a social evaluation process are not well understood. To study these problems, we conducted an experiment where a population of automatic composition models, which simulate a population of human creators, evolves while being evaluated by many listeners. As a result, we found that adaptive evolutions of music styles can occur when blending inheritance of high-dimensional statistics representing composition styles is incorporated in the generation update process. A significant difference in musical preference depending on musical experience was also found. The results suggest a possibility of constructing a system of automatic composition in a new and preferred music style by the experimental framework.

The purpose of this study was to clarify appropriate timing for helping people to join in others’ conversation by quantifying continuous changes in people's willingness to engage in the conversation. As online communication has become more popular in recent years, it is important to elucidate how one can support many people to actively participate in an ongoing conversation. We conducted an experiment to obtain continuous changes in willingness to join in a multi-person discussion conversation presented on a computer monitor. As a result, we found that the willingness to engage in the conversation increased when laughter and silence occur in the discussion. We believe that the results of the present study can be used to develop a dialogue system that supports people to easily join in a conversation by intentionally inserting laughter and silence when people want to express their opinions but are not sure of the timing to speak.

We develop an fair and strategy-proof scheduling mechanism when an agent has an ordering of preference of slot of time. The scheduling problem is known as an application of cake-cutting problem which is to fairly allocate a divisible goods among agents. We formalize a scheduling problem as a cake-cutting problem and then propose an fair and strategy-proof scheduling mechanism. We also evaluate the efficiency obtained the proposed mechanism using computational simulations.

Collective behavior is one of the most fundamental yet challenging phenomena in understanding of animal and even human groups. Previous studies of collective behavior in animal groups have often focused on one-time or short-term performance, largely missing the potential of these systems to learn or to undergo changes over time. To address this problem, we introduced a computational simulation environment based on multi-agent deep reinforcement learning. Here, we studied cooperative hunting, which is a typical example of collective behavior, and found that an individual originally playing a role suddenly changed its role to another at certain time point through learning when two predators cooperated to capture prey. On the other hand, when three predators cooperated, there was no such clear role specialization by individuals and their roles were interchanged more flexibly. Furthermore, we found that the proportion of successful predation can oscillate over time, regardless of role division and its consistency. These results complement existing findings established by observations in nature and provide insight for further understanding of collective animal behavior.

Equation discovery identifies governing equations of dynamics from observations, which is significant for our more profound understanding of the systems. Among equation discovery methods, Sparse Identification of Nonlinear Dynamics (SINDy) has recently attracted considerable attention. SINDy identifies differential equations from the perspective of sparse regression in a high-dimensional nonlinear function space. However, SINDy often contains redundant terms requiring more criteria for selecting variables and functions. To eliminate dull terms based on causality and obtain equations that efficiently describe dynamics, we propose Parsimonious Equation Learning with Causality (PELC). PELC discovers causal networks from multivariate time series via adversarial generative networks and incorporates this topology as a constraint in the hypothesis space of SINDy. We compared the reproducibility of differential equations among SINDy, VAR-LiNGAM, and PELC. As a result, the reproducibility of PELC was the highest. PELC is expected to be a novel method that connects causal network discovery in continuous algebraic space by deep learning and equation discovery.

Online surveys are very useful for planning and verifying policies in many fields such as marketing because of their high cost-effectiveness and ease. However, due to difficulties to conduct it by random sampling, the survey results often contain selection bias. To cope with this problem, the method has been proposed by modeling the occurrence of selection bias and correcting it based on statistical decision theory. To apply this method to analyzing online surveys, it is necessary to put it into a specific model and examine its performance. In this study, we consider correcting selection bias in online surveys in which the response is binary and covariates are represented by continuous values, and assume logistic regression model as a data generation model. Then, we develop a correction method using a selection bias correction framework based on statistical decision theory. We also clarify its properties in numerical experiments on artificial data.

In the field of conventional recommendation systems, most of the models have been based on the prediction of evaluation values using evaluation value data directly assigned by users to their satisfaction with items. Recently, recommendation models that utilize behavioral history data such as implicit evaluation have been widely used. Neural Collaborative Ranking is a method for estimating and ranking the next most likely items to be observed in the list of items. Whereas, there are cases in which multiple implicit evaluations at different levels are observed, such as purchasing and browsing. However, the conventional NCR model cannot distinguish and learn multiple implicit evaluations, and cannot fully utilize the observed data. Therefore, in this study, we propose a model that takes into account multiple implicit evaluations with different levels in the NCR by adopting the method of Ding et al. In addition, we demonstrate the effectiveness of the proposed method.